AI News · Tools
Matt Pocock's /grill-me:
grilling a real pricing decision.
DE Auf Deutsch lesen
/grill-me is meant to sharpen a plan through relentless pushback instead of
nodding it through. /grill-with-docs does the same and writes down a glossary
and architecture decisions along the way. To see what that actually produces, I didn't
simulate a made-up design discussion. I put a real, unresolved pricing question from
PlantWiz in front of it and answered three rounds of questions.
productivity/grill-me,
productivity/grilling and engineering/domain-modeling. Tested in
an isolated Git worktree of PlantWiz, a production Vue 3 application with a Node backend, so
ongoing work in the actual repository stayed untouched. The questions and answers below are
the actual conversation, trimmed of repetition,
the full transcript is here.
What the skills actually prescribe
/grill-me, despite its name, isn't a standalone interview mechanism. The skill
file contains just one line: "Call the Skill tool with 'grilling'." The actual logic lives in
grilling, which, like /handoff from the
earlier post, never loads itself thanks to
disable-model-invocation: true. grilling sets out four rules:
- Design tree instead of a question list: every decision branches into the decisions that hang off it.
- The frontier principle: only ask questions whose prerequisites are already settled. A question whose answer depends on one still open belongs to a later round.
- Facts are the agent's job, decisions are the user's: anything that can be looked up in code or the filesystem should be found by the agent, not asked about.
- Done only when the frontier is empty: the round is finished once every branch of the tree has been visited, not once the first plausible-sounding answer arrives.
/grill-with-docs is even thinner in the original repository than
/grill-me: the skill file simply calls two skills at once, grilling
and domain-modeling. The latter writes, live, the moment a term settles, into
CONTEXT.md (a pure glossary, explicitly free of implementation detail), and files
an ADR under docs/adr/, an Architecture Decision Record:
a short, sequentially numbered document that records which design decision was made and why,
so it doesn't get second-guessed later without context. An ADR only gets created, though, when
a decision meets all three conditions: hard to reverse, surprising without context, and the
result of a real trade-off.
The test case: a real, unresolved pricing question
PlantWiz's LAUNCH-TODO.md lists "finalize pricing" as an open item under Phase 3,
with a placeholder proposal of €4.99/month or €29.99/year. The free/pro boundary already
existed in code: backend/src/utils/planLimits.ts allowed 3 gardens and 1,000
plants per garden for free, a limit almost nobody hits. Not a made-up example, but the actual
state of the repository on the day I triggered the skill.
| Session result | Value |
|---|---|
| Interview rounds | 3 (plus 1 round in the strict frontier format for /grill-with-docs) |
| Questions asked | 10 |
| Starting proposal | Annual plan, €4.99/month or €29.99/year, free limit 1,000 plants |
| End result | Seasonal plan March–October, free limit 100 plants, seasonally gated trial |
| Flagged by the skill itself as its own call | 1 recommendation (beta-user transition) |
Round by round: how the plan changed
The first round put the uncomfortable base question directly: with a free limit of 1,000 plants that nobody hits, what exactly would move a user to pay? On top came follow-ups on where the price numbers came from, the trial length, and how it ties to the beta vouchers from Phase 2. The answer to the pricing question changed the rest of the session: "The problem is nobody needs this app in winter."
A naive annual plan
€4.99/month or €29.99/year, free limit 1,000 plants. The starting proposal from
LAUNCH-TODO.md, taken over unchanged.
Free limit lowered, a contradiction surfaces
Free limit lowered to 100 plants. At the same time it becomes visible: an annual plan for an app that "nobody needs" in winter bakes five unused months into the price and leans on cancellation inertia rather than actual value.
The fix attempt gets dropped
The obvious fix, offsetting the winter problem with a storage-management feature, got dropped in the same answer that raised it: "probably wouldn't help much." The contradiction from step 2 stayed unresolved.
A seasonal plan instead of an annual one
Pressed again on what still justifies the annual price, the actual decision came out: bill Pro only for March–October, start the trial only outside winter. An answer that wasn't in the original proposal at all.
grilling actually prescribes, "the decisions belong to the user, never the
agent" is written into the skill file word for word. When the user hands the decision back,
it becomes an agent decision in practice after all, just named openly instead of made
silently.
The difference with /grill-with-docs
The first three rounds ran with plain /grill-me: questions, answers, decisions,
none of it lands in a file. For the last step, two remaining detail questions on validating
the free limit and the beta transition, I additionally installed grill-with-docs
and domain-modeling and continued the same session with them. The moment the
frontier was empty, two files appeared unprompted:
# PlantWiz
Web app for garden bed planning for hobby gardeners.
## Language
**Seasonal Plan**:
Pro subscription billed only for the March–October period, because the app sees
almost no use outside the growing season.
_Avoid_: Annual plan, standing subscription
**Free Limit**:
The usage ceiling in the free plan, past which an upgrade to Pro is needed.
_Avoid_: Limit, quotaand, because the decision met all three ADR criteria (hard to reverse, surprising without context, the result of a real trade-off):
# Seasonal plan instead of annual plan for Pro
PlantWiz sees almost no use outside the growing season. A classic annual plan
bakes in five months in which the user gets no value, leaning on cancellation
inertia rather than actual use. A checked proposal to offset the winter months
with a storage-management feature was dropped. Decision: Pro is billed only for
March–October; the trial also starts only within that window.Lasting documentation after the session
Same decision, same session, only the last step differs.
with-docs
509 characters of glossary, 664 characters of ADR. No compression win like with
/handoff, and that isn't the point here: without domain-modeling,
the seasonal-plan decision lives only in this session's chat history, invisible to the next
look at the repository. With it, two short files carry why PlantWiz no longer sells an
annual plan, for anyone who opens the repository afterward, human or agent.
Direct comparison to the obvious alternatives
| /grill-me | /grill-with-docs | Deciding directly, no interview | |
|---|---|---|---|
| Surfaces contradictions in the plan | yes, through targeted questions | yes, plus checking against the existing glossary | no, only what you happen to notice |
| Leaves lasting documentation | no | yes, CONTEXT.md and ADRs | no, unless you write it yourself |
| Forces an order (frontier) | yes, later questions wait on earlier answers | yes, plus a terminology check | no |
| Speed | several rounds, depending on the topic's depth | same, plus time to write the docs | fastest |
| Risk | agent recommendation on decisions handed back | same risk, plus a chance of superfluous ADRs from a too-generous reading of the three criteria | blind spots stay blind |
When it's worth it
A plan that feels too smooth
When a proposal like "€4.99/month, €29.99/year" gets settled quickly, but nobody has checked the numbers against actual usage. Exactly the case from this test.
The decision needs to stay findable in the code
/grill-with-docs instead of /grill-me, once the decision is
hard to reverse. A price you change again after launch upsets existing customers; that
belongs in an ADR, not only in the chat history.
A pure implementation question with no real alternative
Skip /grill-me. When there's only one reasonable path, the interview
mostly produces questions that answer themselves, and just costs time.
/grill-with-docs preserves the result
in two files that would have been lost after the session with plain /grill-me
unless someone wrote them down by hand.
Choosing skills for your team?
I set up Claude Code in developer teams, including the question of which skills actually pay off and which just cost tokens. 30-minute intro call, free of charge.